feat: M3 — 伏笔账本 + 节奏引擎 + 大纲(含并发记账 bugfix)

- 伏笔账本:纯函数状态机(OPEN/PARTIAL/CLOSED/OVERDUE) + ForeshadowLedger repo;验收后到期扫描(BackgroundTask 自建 session 置 OVERDUE);登记/状态变更端点
- 节奏 + 三审齐:foreshadow-analyst + pace-checker 并入 LangGraph 并行审(REVIEW_SPECS),collect 分列落 chapter_reviews(conflicts/foreshadow_sug/pace),review SSE 加 foreshadow/pace 事件
- 大纲:outliner Agent 产 OutlineResult(含 foreshadow_windows),POST /outline 逐章 upsert outline 表;GET /foreshadow?status= 看板
- 前端:伏笔四泳道看板(OVERDUE 琥珀) + 大纲编辑器(窗口徽标) + 节奏节拍图(▁▃▅) + 审稿页消费 foreshadow/pace SSE
- bugfix(T3.8):并行三审共用请求 session 记账触发 'Session is already flushing' → foreshadow/pace 静默丢失;SqlAlchemyLedgerSink.record 改 add-only(靠端点/事务 commit),加并发回归测试
- M3 E2E:真实 DB + mock 网关零 token 走通 埋设→进展→验收后扫描 OVERDUE→看板 + 大纲含窗口 + 三审齐 SSE/留痕;E2E 暴露并钉住上述 bug
- 门禁绿:mypy 111 / pytest 228(0 xfailed) / alembic 无漂移;前端 gen:api/lint/tsc/vitest 69/build
This commit is contained in:
Yaojia Wang
2026-06-18 14:21:17 +02:00
parent 68f194a043
commit 5fb7bfb1de
74 changed files with 6529 additions and 126 deletions

View File

@@ -0,0 +1,112 @@
"""T3.8 回归并行审记账并发安全add-only无 flush 重入)。
旧 bugT3.7 真 pg E2E 暴露):三审作为同一 LangGraph superstep 的并行分支,各自
`gateway.run()` → 共用**同一请求 session** 的 `SqlAlchemyLedgerSink.record`
(`session.add` + `await session.flush()`)。`AsyncSession` 非并发安全:`await flush()`
是并行路径里唯一让出控制权的 DB-IO第二/三审的 flush 撞上「Session is already
flushing」flush 重入)→ 被 `run_review` 失败隔离吞成 `incomplete` → foreshadow/pace
结果静默丢失。
本测试用一个会**侦测 flush 重入**的假 AsyncSession`flush` 异步(`await asyncio.sleep`
让出控制权),若两协程同时在 flush 中即抛 `RuntimeError("Session is already flushing")`
——精确复刻 SQLAlchemy 的报错。在旧 add-then-flush 实现下,`asyncio.gather` 三并发
`record` 会触发该重入;修复后 `record` 只 `add`(同步、不让出),不再调 flush故三行
全部入 session 且零重入。
"""
from __future__ import annotations
import asyncio
import uuid
import pytest
from ww_llm_gateway.ledger import SqlAlchemyLedgerSink
from ww_llm_gateway.types import Scope, Usage
class _FlushReentryDetectingSession:
"""假 AsyncSession记录 `add` 的行;`flush` 异步且侦测并发重入。
`add` 同步追加(镜像真 session。`flush` 进入时若已有协程在 flush 中 → 抛
`RuntimeError("Session is already flushing")`(复刻 SQLAlchemy 行为),否则
`await asyncio.sleep(0)` 让出控制权模拟真实 DB-IO 让步点。
"""
def __init__(self) -> None:
self.added: list[object] = []
self.flush_calls = 0
self._flushing = False
def add(self, row: object) -> None:
self.added.append(row)
async def flush(self) -> None:
self.flush_calls += 1
if self._flushing:
raise RuntimeError("Session is already flushing")
self._flushing = True
try:
await asyncio.sleep(0) # 让出控制权——并行协程在此交错
finally:
self._flushing = False
def _usage(input_tokens: int) -> Usage:
return Usage(
provider="deepseek",
model="deepseek-chat",
input_tokens=input_tokens,
output_tokens=1,
cost_minor=0,
currency="USD",
)
def _scope() -> Scope:
return Scope(user_id=uuid.UUID(int=1), project_id=uuid.UUID(int=2))
async def test_concurrent_record_adds_all_rows_without_flush_reentry() -> None:
"""三并发 `record`(模拟三审并行)→ 三行全部入 session绝不抛 flush 重入。
旧 add-then-flush 实现在此会因 `await flush()` 让出控制权而 flush 重入抛错;
add-only 实现下 `record` 不让出控制权,三行确定性入 session。
"""
session = _FlushReentryDetectingSession()
sink = SqlAlchemyLedgerSink(session) # type: ignore[arg-type]
await asyncio.gather(
sink.record(_scope(), _usage(100)),
sink.record(_scope(), _usage(200)),
sink.record(_scope(), _usage(300)),
)
# 三审用量全部入 session一次调用一行记账不丢
assert len(session.added) == 3
input_tokens = sorted(row.input_tokens for row in session.added) # type: ignore[attr-defined]
assert input_tokens == [100, 200, 300]
# add-onlysink 不在并行路径 flush持久化靠端点/事务 commit
assert session.flush_calls == 0
async def test_record_does_not_flush_so_caller_commit_persists() -> None:
"""单次 `record` 只 add 不 flush——契约「网关 ledger 只 add、调用方负责 commit」。"""
session = _FlushReentryDetectingSession()
sink = SqlAlchemyLedgerSink(session) # type: ignore[arg-type]
await sink.record(_scope(), _usage(42))
assert len(session.added) == 1
assert session.flush_calls == 0
@pytest.mark.parametrize("n_concurrent", [2, 5, 10])
async def test_concurrent_record_scales_without_loss(n_concurrent: int) -> None:
"""N 并发 `record` 全部落 session无丢失、无 flush 重入(不论并发度)。"""
session = _FlushReentryDetectingSession()
sink = SqlAlchemyLedgerSink(session) # type: ignore[arg-type]
await asyncio.gather(*(sink.record(_scope(), _usage(i + 1)) for i in range(n_concurrent)))
assert len(session.added) == n_concurrent
assert session.flush_calls == 0